Privacy-Aware EMG signal synthesis for ergonomic studies using a generative diffusion model

Electromyography (EMG) is an important tool for monitoring muscle activity level during occupational tasks and thereby supporting preventive measures to mitigate work-related musculoskeletal disorders (WMSDs). Machine learning techniques have recently advanced EMG data analysis, enabling improvements in occupational activity recognition and ergonomics assessments. However, these techniques face challenges from limited EMG data availability due to privacy constraints. To address this, our research introduces a U-Net-based diffusion model for synthesizing EMG signals associated with common manual material handling (MMH) tasks. This model aims to preserve the statistical characteristics of original EMG data in the synthesized signals. Evaluation shows that the synthesized data visually resembles the original in both time and frequency domains, capturing similar patterns. Quantitative analysis further demonstrates that the model effectively captures essential EMG features, with statistical testing showing no significant evidence that the original and synthesized data differ in distribution. A classifier trained on real EMG data to identify four types of occupational tasks performs reliably when tested on synthesized EMG signals, demonstrating the feasibility of using EMG data synthesis for activity classification tasks. These findings highlight the effectiveness of the proposed EMG data synthesis method, which addresses privacy concerns and enriches EMG datasets for biomedical and occupational health research.